| Despite their importance in ecosystems and climate change, there are few machine learning models on the ocean carbon cycle and biological carbon pump. This project explores the possibility of applying deep learning techniques to ocean modeling by using two types of machine learning models: a spatiotemporal neural network with convolutional and long-short-term memory (LSTM) layers and a graph neural network (GNN). With the past observations of ocean variables including dissolved oxygen, temperature, salinity, etc., our models can predict the value of net community productions (NCP), a pivotal parameter to evaluate the biological carbon pump. Our models can effectively capture the basic pattern of NCP on both 2-dimensional and 3-dimensional datasets. Of the two models, the GNN model shows better robustness and flexibility, indicating its potential in ocean modeling. I hope that this study brings novel insights into ocean modeling with deep learning, benefitting both earth science and data science communities. However, there are several issues in our models, the serious of which is numerical vulnerability. I hope to learn more about oceanography and advanced deep learning techniques to optimize our models comparable to the state-of-the-art deep learning models in earth science. |
